Claude Code Tutorial: Commands, Modes, and Memory

Step 0 — Install and authenticate

Two install paths. The native installer is the current recommendation; npm works if you already run Node 18+:

# macOS / Linux
curl -fsSL https://claude.ai/install.sh | bash
# Windows (PowerShell)
irm https://claude.ai/install.ps1 | iex
# npm alternative
npm install -g @anthropic-ai/claude-code
claude --version

Then claude from any project directory. The first run opens an OAuth browser flow — a Claude Pro seat ($20/mo) is enough to start; Max tiers ($100–200/mo) buy more headroom when you hit usage limits. API-key auth works too (claude setup-token generates a long-lived token for CI). Verify with claude --help.

Desktop app or VS Code extension instead of the terminal? They share your account and your project's CLAUDE.md — start wherever you are comfortable and move later. This tutorial covers the CLI, because everything else layers on top of it.

Step 1 — First session: the agent loop

Start in a real project — that is where Claude Code differs from a chat window:

cd my-project
claude
> Explain the architecture of this project. Do not modify anything.

Watch the loop: Claude picks a tool (Glob to scan, Read to open, Grep for patterns), executes it, reads the result, and decides whether it needs another pass. Every tool call prints in your terminal. This loop — goal, tool, result, repeat — is the entire mental model. Everything below just controls it.

Three input prefixes matter from minute one: @src/auth.ts attaches a file to your prompt, !ls -la runs a shell command inline (output goes to Claude), and # adds a fact to project memory. Esc interrupts mid-turn and keeps work done so far; Shift+Tab cycles permission modes (next section).

Step 2 — The commands that matter

Claude Code ships dozens of commands; ten cover about 90% of real sessions. A command is only recognized at the start of your message, everything after the name becomes its arguments, and /help lists what your version actually exposes — versions move fast, so type / and filter rather than memorizing.

CLI flags round out the session layer: claude -p "query" is print mode for scripts and CI (pipe-friendly, exits when done), claude --permission-mode plan boots read-only, and cat file | claude -p "summarize" is the headless pattern. Everything else — --verbose, --add-dir, claude mcp for server config — you will pick up when you need it.

Step 3 — Permission modes: the control surface

This is the part most tutorials wave past and the part that decides whether you trust the agent. Shift+Tab cycles modes; the current one shows in the status bar:

The pattern that works in practice: plan → edit → verify. Plan the approach read-only, switch modes and let it execute, then run the test suite yourself. Skip the plan phase on typo-level tasks — that round trip is pure overhead there.

Step 4 — CLAUDE.md: the memory that makes it good

Claude Code is stateless between sessions until you give it a memory file. Run /init in your repo and it generates a starting CLAUDE.md from your codebase (it even imports existing Cursor rules, Copilot instructions, and AGENTS.md). The generated file is a first draft — the highest-leverage edit is making it short and specific:

# Project: billing-service
## Stack
- Bun + Hono on PostgreSQL (Drizzle ORM). Never suggest npm.
## Commands
- bun dev / bun test / bun db:migrate
## Conventions
- TypeScript strict. Tests colocated as *.test.ts.
## Don't
- Don't touch src/legacy/ — frozen.
- Don't add dependencies without asking.

Keep it under ~800 tokens of facts Claude needs every session — build commands, conventions, boundaries. Multi-step procedures belong in skills (next section), not here. The three sections that earn their tokens: the stack block (kills wrong-framework suggestions), the Don't block (your most common agent mistakes from past sessions), and the commands block (so it never guesses variants).

Scope layers naturally: ~/.claude/CLAUDE.md is your global default, ./CLAUDE.md is team-shared project memory, subdirectory files apply when Claude works in them (nearest wins on conflict), and CLAUDE.local.md holds private per-project preferences you gitignore. Auto memory is the other half: Claude writes its own notes across sessions (debugging insights, style preferences) into ~/.claude/projects/<project>/memory/ — toggle it with /memory. Our AGENTS.md template covers the vendor-neutral twin of this file if your team runs mixed agents.

Step 5 — Skills: your own slash commands

Anything you paste into chat more than twice wants to be a skill. A skill is a folder with a SKILL.md (YAML frontmatter + instructions); its directory name becomes the command. Drop it in .claude/skills/deploy/SKILL.md and /deploy exists — for you, for teammates who pull the repo, and loadable by Claude automatically when relevant:

---
name: deploy
description: Ship the current branch to staging with checks.
disable-model-invocation: true
---
Run the pre-deploy checklist: bun test, bun lint, then
deploy to staging with bun run deploy:staging. Report
each gate's result before moving to the next.

The two frontmatter flags that matter: disable-model-invocation: true makes it you-only (right for anything with side effects like deploys — Claude should never decide to deploy), and context: fork runs it in an isolated subagent. Long reference material is nearly free in a skill — the body loads only when used — while CLAUDE.md content taxes every message. That split is the whole design.

Old .claude/commands/*.md files still work — custom commands merged into skills. MCP servers extend tools instead of commands: wire them with claude mcp, and see our MCP tools guide for the servers worth attaching.

Step 6 — Verify (the step beginners skip)

The agent says done. Your job is not done. Three closes, every time:

This is not ceremony — it is the loop that separates developers who get real leverage from agents from those who ship subtle bugs. Our vibe coding security guide covers the failure modes (45% of AI-generated code carries an OWASP flaw in 2026 studies) and the 8-step review loop that closes them.

Run Claude Code beside your other agents

Claude Code is a terminal agent; a harness decides where it runs and what it can touch. VibeFuse is the first ever free widget-based AI harness: Claude Code, Codex, and Gemini as live widgets on one Windows canvas — sessions named and resumable, MCP tools attached, local Whisper voice control, all processing locally. The harness is $0; you pay only the model vendors you already use.

Skills you write for Claude Code are portable: VibeFuse reads the same SKILL.md format, and the open-source marketplace lets you sell your skills, widgets, and styling packs with 80% payouts to creators. Works in any app you already use.

Explore VibeFuse & harness guides

• Harness Guide
• Free Coding Tools
• AI Coding Agent Desktop
• Free Voice Transcription
• Free Text to Speech
• VibeFuse Product
• Widget Marketplace
• Download Free
• VibeFuse Docs
• Shareable AI Widgets
• Shareable AI Skills
• MCP Tools
• AI Agent Harness
• Harness Engineering
• HyperFrames Video
• Skill Seekers
• Sell AI Skills
• AI Skills Marketplace Compared
• Cursor Alternative
• Local Whisper + Piper
• Custom AI Dashboard
• Community Hub